`sts_zoo.animals`

`LLTFactory`

A local linear trend (LLT) factory.

The generative model for programs created by this factory is

\[\begin{aligned} l_t &= l_{t - 1} + \mathrm{loc}_t + \mathrm{scale\_local}_t w^l_t,\ l_0 = \mathrm{ic\_local} \\ z_t &= z_{t - 1} + l_t + \mathrm{scale\_level}_t w^z_t,\ z_0 = \mathrm{ic\_level}, \end{aligned}\]

where \(w^\cdot_t \sim \mathrm{Normal}(0, 1)\), \(t_0 \leq t \leq t_1\).

Args:

loc (Block || torch.tensor || pyro.distributions): location parameter
scale_local (Block || torch.tensor || pyro.distributions): scale parameter for \(l_t\)
scale_level (Block || torch.tensor || pyro.distributions): scale parameter for \(z_t\)
ic_local (torch.tensor || pyro.distributions): initial condition for \(l_t\)
ic_level (torch.tensor || pyro.distributions): initial condition for \(z_t\)

`make`

Creates a new LLT model. The sites in the created model are guaranteed to be unique.

Args:

None

`ModelFactory`

Abstract base class for all model factories. Ensures unique ids for each component of each generated model, and encapsulates model generation through make method.

`make`

Abstract method to be overridden. Calling .make() should create a new stsb3 model with guaranteed unique site names.

`SGTFactory`

A seasonal + global trend (SGT) factory.

The generative model for programs created by this factory is

\[\begin{aligned} l_t &= \alpha_t + \beta_t l_{t-1} + \mathrm{scale\_local}_t w_t \\ g_t &= a + bt \\ s_t &= \theta_{t \mathrm{mod} s},\ s=1,...,\mathrm{num\_seasons}\\ z_t &= s_t + g_t + l_t, \end{aligned}\]

where \(w_t \sim \mathrm{Normal}(0, 1)\), \(t_0 \leq t \leq t_1\), and \(\mathrm{num\_seasons}\) is the number of discrete seasons.

Args:

alpha (Block || torch.tensor || pyro.distributions): intercept parameter of \(l_t\)
beta (Block || torch.tensor || pyro.distributions): slope parameter of \(l_t\)
a (Block || torch.tensor || pyro.distributions) intercept parameter of \(g_t\)
b (Block || torch.tensor || pyro.distributions) slope parameter of \(g_t\)
num_seasons (int): \(\geq 2\) integer number of seasons, defaults to 2
seasons (torch.tensor || pyro.distributions) season values
scale_local (Block || torch.tensor || pyro.distributions): scale parameter for \(l_t\)
scale_level (Block || torch.tensor || pyro.distributions): scale parameter for \(z_t\)
ic_level (torch.tensor || pyro.distributions): initial condition for \(z_t\)

`make`

Creates a new SGT model. The sites in the created model are guaranteed to be unique.

Args:

None

`SLLTFactory`

A semi-local linear trend (SLLT) factory.

The generative model for programs created by this factory is

\[\begin{aligned} l_t &= \alpha_t + \beta_t l_{t-1} + \mathrm{scale\_local}_t w^l_t \\ z_t &= z_{t - 1} + l_t + \mathrm{scale\_level}_t w^z_t,\ z_0 = \mathrm{ic\_level}, \end{aligned}\]

where \(w^\cdot_t \sim \mathrm{Normal}(0, 1)\), \(t_0 \leq t \leq t_1\).

Args:

alpha (Block || torch.tensor || pyro.distributions): intercept parameter of \(l_t\)
beta (Block || torch.tensor || pyro.distributions): slope parameter of \(l_t\)
scale_local (Block || torch.tensor || pyro.distributions): scale parameter for \(l_t\)
scale_level (Block || torch.tensor || pyro.distributions): scale parameter for \(z_t\)
ic_level (torch.tensor || pyro.distributions): initial condition for \(z_t\)

`make`

Creates a new SLLT model. The sites in the created model are guaranteed to be unique.

Args:

None

`StochVolFactory`

A stochastic volatility model factory.

This is a model for the log of an asset price \(S_t\). Define \(z_t = \log S_t\). The generative model for programs created by this factory is

\[\begin{aligned} \mathrm{isp}\ \sigma_t &= \mathrm{isp}\ \sigma_{t - 1} + \mathrm{d}t_t\ \mathrm{loc\_ispvol}_t\ + \sqrt{\mathrm{d}t_t}\ \mathrm{scale\_ispvol}_t\ w^{\mathrm{isp}\sigma}_t,\ l_0 = \mathrm{ic\_logvol} \\ z_t &= z_{t - 1} + \mathrm{d}t_t\mathrm{loc}_t + \sqrt{\mathrm{d}t_t}\sigma_t w^z_t,\ z_0 = \mathrm{ic}, \end{aligned}\]

where \(w^\cdot_t \sim \mathrm{Normal}(0, 1)\), \(t_0 \leq t \leq t_1\), and \(\mathrm{isp}\) is the inverse softplus function, \(\mathrm{isp}(y) = \log(-1 + \exp y)\).

Args:

dt (Block || torch.tensor || pyro.distributions): time discretization. If a pyro.distributions object, it should be an atomless distribution with non-negative support and will be interpreted as \(p(dt) = \prod_{t_0\leq t\leq t_1} p(dt_t)\), not as a single draw. Passing a distributions object can be a useful method for modeling jump-y asset prices (e.g., \(p(dt_t) = \mathrm{Gamma}(dt_t | \alpha, \beta)\) generates the variance gamma process). However, doing this also makes the model non-identifiable in the scale parameter(s).
loc_ispvol (Block || torch.tensor || pyro.distributions): location parameter of the isp volatility process
vol_ispvol (Block || torch.tensor || pyro.distributions): scale/volatility parameter of the isp volatility process
ic_ispvol (torch.tensor || pyro.distributions): initial condition of the isp_volatility process
scale_level (Block || torch.tensor || pyro.distributions): scale parameter for \(z_t\)
loc (Block || torch.tensor || pyro.distributions): location parameter for the log asset price process.
ic (torch.tensor || pyro.distributions): initial condition for the log asset price process

`make`

Creates a new StochVol model. The sites in the created model are guaranteed to be unique.

Args:

None